Oracle Analysis

ORACLE_ANALYSIS Skill

Trigger Pattern: ORACLE flag (required) Inject Into: Breadth agents, depth-external, depth-edge-case Purpose: Analyze all oracle integrations in Aptos Move protocols for staleness, decimal errors, zero/negative prices, confidence intervals, multi-oracle aggregation, and failure modes

For every oracle the protocol consumes:

STEP PRIORITY: Steps 6 (Failure Modes) and 5c (Deviation Reference) are where HIGH/CRITICAL severity findings most commonly hide. Do NOT rush these steps. If constrained, skip conditional sections (4a-4d, 5a) before skipping 5c or 6.

1. Oracle Inventory

Enumerate ALL oracle data sources the protocol reads:

Aptos oracle landscape:

• Pyth Network: pyth::price_feed module, returns Price { price: I64, conf: u64, expo: I64, publish_time: u64 }
• Switchboard: switchboard::aggregator module, returns aggregator results with mantissa and scale
• Custom price feeds: Protocol-specific oracles using Table or SmartTable for price storage
• On-chain TWAP: DEX-derived time-weighted prices (Thala, LiquidSwap, Pontem)

For each oracle: What decision does the protocol make based on this data? (pricing, liquidation threshold, reward rate, rebase trigger, collateral valuation, etc.)

Hardcoded stablecoin pricing: Does the protocol skip oracle lookup for any asset and hardcode its price (e.g., USDC = 1e8)? All assets require dynamic oracle pricing — stablecoins depeg.

2. Staleness Analysis

For each oracle identified in Step 1:

2a. Staleness Checks Present?

Pyth-specific: Is price.publish_time compared against timestamp::now_seconds()? What max age is enforced? Switchboard-specific: Is the aggregator's latest_confirmed_round.round_open_timestamp validated?

Chained feed deviation: If derived prices require multiple feeds (e.g., token_A/USD via token_A/APT + APT/USD), sum individual deviation thresholds to compute total worst-case deviation. If total exceeds LTV buffer → FINDING.

If NO staleness check: What happens when the oracle returns stale data?

• Protocol uses stale price for liquidations -- unfair liquidations
• Protocol uses stale price for minting -- mispriced assets
• Protocol uses stale price for swaps -- arbitrage opportunity
• Protocol uses stale rate for rewards -- incorrect distribution

2b. Stale Data Impact Trace

For each consumer function, trace the impact of receiving data that is {freshness_guarantee x 2} old:

2c. Pyth-Specific Checks

2d. Switchboard-Specific Checks

3. Decimal Normalization Audit

For each oracle data flow:

Pyth decimal handling: Pyth uses expo field (typically negative, e.g., expo = -8 means 8 decimal places). The actual price = price.price * 10^expo. Common errors:

• Treating expo as positive when it is negative
• Not converting I64 exponent to unsigned for power calculation
• Mixing Pyth's expo-based decimals with token decimals (Aptos Coin typically uses 8 decimals, but FungibleAsset varies)

Switchboard decimal handling: Uses mantissa and scale (or decimals). Actual value = mantissa * 10^(-scale).

MANDATORY GREP: Search all oracle consumer files for hardcoded decimal constants: 100000000, 1e8, 10_000_000, DECIMAL, PRECISION. For each hit: (1) Is this a decimal normalization constant? (2) Does it match the ACTUAL oracle's decimal format? (3) If the oracle feed changes or is swapped, does this constant break?

Decimal chain trace: For each arithmetic operation using oracle data, trace the full decimal chain: oracle_output_decimals -> normalization_step -> consumer_expected_decimals. If any step uses a hardcoded constant rather than reading decimals dynamically -> FINDING.

Common decimal mismatches on Aptos:

• Pyth USD feeds: expo = -8 (8 decimals), but protocol assumes 18
• Aptos native Coin: typically 8 decimals
• FungibleAsset: varies per metadata configuration
• Cross-multiplication without normalization: price * amount where price and amount have different decimal bases

3d. Decimal Grep Sweep (MECHANICAL -- MANDATORY)

Grep ALL oracle consumer files for 10_u128|pow\(10|DECIMALS|PRECISION|100000000|normalize. For each match, fill:

If ANY row shows Match=NO or oracle decimals UNKNOWN with hardcoded constant -> FINDING (R16). Skipping this step is a Step Execution violation (x3d).

4. TWAP-Specific Analysis

If protocol uses any TWAP oracle (DEX-derived, custom accumulator, etc.):

4a. TWAP Window Analysis

Rule of thumb: TWAP window < 30 min AND pool TVL < $10M -> potentially manipulable.

4b. TWAP Arithmetic

4c. TWAP Lagging Behavior

During rapid price movements, TWAP lags spot price. Trace:

• What happens when TWAP price is significantly lower than spot? (discounted minting/borrowing)
• What happens when TWAP price is significantly higher than spot? (premium liquidations)
• Is this lag exploitable by attackers who can predict the direction?

4d. TWAP Cold-Start Analysis

Check oracle behavior when history is insufficient: (1) zero snapshots, (2) single snapshot, (3) window period not yet elapsed.

For each exploitable state: can attacker act during cold-start window at manipulated price? Tag: [BOUNDARY:snapshots=0], [BOUNDARY:snapshots=1]. If TWAP returns 0 or aborts during cold-start with no fallback -> FINDING (R16, minimum Medium).

5. Oracle Weight / Threshold Boundaries

For multi-oracle systems or oracle-based thresholds:

5a. Multi-Oracle Systems

Check: What happens at exact threshold boundaries?

• If median of [100, 100, 101]: result = 100. Is that correct?
• If weighted average with equal weights rounds down: impact?
• If one oracle call aborts: does fallback handle it gracefully?

5b. Oracle-Based Thresholds

Check > vs >=: At the exact threshold value, does the protocol behave as intended?

5c. Deviation Reference Point Audit

For each deviation check in the protocol (maxDeviation, priceDeviation, deviationThreshold, etc.):

Checks:

1. What is the deviation MEASURED AGAINST? (previous on-chain price, TWAP, external oracle, hardcoded value)
2. Is the reference point itself manipulable? (e.g., if deviation checks current vs last-recorded, and last-recorded is admin-settable -> admin can set a stale reference that makes all future prices "within deviation")
3. Can the reference become stale? (e.g., if reference is updated only on specific actions, and those actions stop occurring)
4. Is the first recorded price special? (no prior reference -> deviation check may be bypassed on first update) Tag: [TRACE:deviation check: current vs {reference} -> reference source: {X} -> manipulable: {Y/N}]

6. Oracle Failure Modes

For each oracle, model failure scenarios:

Aptos-specific failure note: Move does not have try/catch. Oracle call failures result in transaction abort. This means:

• External oracle call that aborts -> entire transaction reverts
• No graceful fallback unless protocol pre-checks oracle state with exists<> or similar
• Oracle DoS (feed stops updating) -> all dependent functions become uncallable

For each unmitigated failure mode: What is the worst-case impact? Can it lead to fund loss?

Circuit breaker check: Does the protocol have a mechanism to pause oracle-dependent operations if the oracle enters a failure state?

Instantiation Parameters

{CONTRACTS}           -- Move modules to analyze
{ORACLE_MODULES}      -- Oracle module paths (pyth::price_feed, switchboard::aggregator, custom)
{CONSUMER_FUNCTIONS}  -- Functions that read oracle data
{PRICE_FEED_IDS}      -- Pyth price feed identifiers or Switchboard aggregator addresses
{TOKEN_DECIMALS}      -- Decimal configuration of tokens in scope

Finding Template

**ID**: [OR-N]
**Severity**: [based on fund impact and likelihood of oracle failure/manipulation]
**Step Execution**: checkmark1,2,3,4,5,6 | x(reasons) | ?(uncertain)
**Rules Applied**: [R1:Y, R4:Y, R10:Y, R16:Y]
**Location**: module::function:LineN
**Title**: Oracle [issue type] in [function] enables [attack/failure]
**Description**: [Specific oracle issue with data flow trace]
**Impact**: [Quantified impact under worst-case oracle scenario]

Output Schema

Step Execution Checklist (MANDATORY)